def choose_action(self, s, visual_s, evaluation=False):
if self.use_epsilon and np.random.uniform(
) < self.expl_expt_mng.get_esp(self.episode, evaluation=evaluation):
a = np.random.randint(0, self.a_counts, len(s))
else:
a = self._get_action(s, visual_s)[-1].numpy()
return sth.int2action_index(a, self.a_dim_or_list)
def choose_inference_action(self, s, visual_s):
a = self.sess.run(self.action, feed_dict={
self.pl_visual_s: visual_s,
self.pl_s: s,
self.sigma_offset: np.full(self.a_counts, 0.01)
})
return a if self.action_type == 'continuous' else sth.int2action_index(a, self.a_dim_or_list)
def choose_inference_action(self, s, visual_s):
return sth.int2action_index(
self.sess.run(self.action,
feed_dict={
self.pl_visual_s: visual_s,
self.pl_s: s
}), self.a_dim_or_list)
def step(self, actions):
actions = deepcopy(actions)
for i, k in enumerate(actions.keys()):
if self.is_continuous[i]:
pass
else:
actions[k] = sth.int2action_index(actions[k], self.discrete_action_dim_list[i])
return self._env.step(actions)
def choose_action(self, s, visual_s):
if self.action_type == 'continuous':
return self._get_action(s, visual_s).numpy()
else:
if np.random.uniform() < self.epsilon:
a = np.random.randint(0, self.a_counts, len(s))
else:
a = self._get_action(s, visual_s).numpy()
return sth.int2action_index(a, self.a_dim_or_list)
def choose_action(self, s, visual_s):
if np.random.uniform() < self.epsilon:
a = np.random.randint(0, self.a_counts, len(s))
else:
a = self.sess.run(self.action, feed_dict={
self.pl_visual_s: visual_s,
self.pl_s: s
})
return sth.int2action_index(a, self.a_dim_or_list)
def choose_action(self, s, visual_s, evaluation=False):
self._check_agents(s)
a = [
model(s_).numpy() for model, s_ in zip(
self.cem_models, np.split(s, self.populations, axis=0))
]
if self.is_continuous:
a = np.vstack(a)
else:
a = np.hstack(a)
a = sth.int2action_index(a, self.a_dim_or_list)
return a
def choose_action(self, s, visual_s):
if self.action_type == 'continuous':
return self.sess.run(self.action, feed_dict={
self.pl_visual_s: visual_s,
self.pl_s: s,
self.sigma_offset: np.full(self.a_counts, 0.01)
})
else:
if np.random.uniform() < self.epsilon:
a = np.random.randint(0, self.a_counts, len(s))
else:
a = self.sess.run(self.action, feed_dict={
self.pl_visual_s: visual_s,
self.pl_s: s
})
return sth.int2action_index(a, self.a_dim_or_list)
def step(self, actions):
actions = np.array(actions)
if not self.is_continuous:
actions = sth.int2action_index(actions, self.a_dim_or_list)
if self.action_type == 'discrete':
actions = actions.reshape(-1,)
elif self.action_type == 'Tuple(Discrete)':
actions = actions.reshape(self.n, -1).tolist()
results = Asyn.op_func(self.envs, Asyn.OP.STEP, actions)
obs, reward, done, info = [np.asarray(e) for e in zip(*results)]
reward = reward.astype('float32')
dones_index = np.where(done)[0]
if dones_index.shape[0] > 0:
correct_new_obs = self.partial_reset(obs, dones_index)
else:
correct_new_obs = obs
if self.obs_type == 'visual':
obs = obs[:, np.newaxis, ...]
correct_new_obs = correct_new_obs[:, np.newaxis, ...]
return obs, reward, done, info, correct_new_obs
def choose_action(self, adj, x, visual_s, evaluation=False):
a = self._get_action(adj, x, visual_s, evaluation).numpy()
return a if self.is_continuous else sth.int2action_index(a, self.a_dim_or_list)
def choose_inference_action(self, s, visual_s):
a = self._get_action(s, visual_s)[0].numpy()
return a if self.action_type == 'continuous' else sth.int2action_index(
a, self.a_dim_or_list)
def choose_inference_action(self, s, visual_s):
return sth.int2action_index(
self._get_action(s, visual_s).numpy(), self.a_dim_or_list)
def choose_action(self, s, visual_s):
if self.action_type == 'continuous':
return self._get_action(s, visual_s).numpy()
else:
a = self._get_action(s, visual_s).numpy()
return sth.int2action_index(a, self.a_dim_or_list)
def choose_action(self, s, visual_s):
if self.use_epsilon and np.random.uniform() < self.epsilon:
a = np.random.randint(0, self.a_counts, len(s))
else:
a = self._get_action(s, visual_s)[-1].numpy()
return sth.int2action_index(a, self.a_dim_or_list)